A map of AI agents
Lesson 19 of 26 in the course “Vibe coding”. Free, with animations and a quiz.
An agent: AI that takes action
A regular AI chat gives you code, and you copy and paste it yourself. An agent is AI that acts inside your project. It reads files, edits many of them at once, runs terminal commands and tests (programs that check your code), and can make a commit and open a PR. You set the task and check the result; it does the hands-on work. The diagram shows its work loop, and we'll go through it in two slides.
Three kinds of agents
Agents in the terminal or editor: Claude Code, Codex, agent mode in Cursor, the GitHub Copilot agent. They work on the project right on your computer. Cloud agents work in a copy of your repository on the service's servers and hand back a finished PR. Agents inside builders (Lovable, Bolt, v0, Replit Agent) build the whole app in the browser. The diagram shows the columns from the lesson on tools.
The loop: plan, code, check
Example: you ask the agent to add a gig schedule page to your band's site. Plan: the agent reads the project and writes what it will do: a new page, a link in the menu. Code: it edits the files. Check: it runs the site and the tests. Fix: it sees an error and fixes it itself. The chip on the loop in the diagram is a commit, a save after a working step. Then round again until the task is done.
You approve the commands
The little person in the diagram is you, at the “approve command” point. Before running a command, the agent usually asks for permission. Don't hit “yes” without looking. Be extra careful with deleting files, installing unfamiliar packages (other people's code libraries), pushing to main, and anything involving keys. Don't understand a command? Ask the agent what it does. Save the “allow everything” mode for later.
Rules for working with an agent
One task at a time. Ask for a plan first and read it; code comes after. Have the agent work in a separate branch and open a PR, as in the diagram, so the live site stays safe. Commit after every working step. Which agent to start with? Whichever you can access: check that the service and payment work in your country. The skill of setting tasks and checking results carries over to any of them.
You're still the producer
An agent does more than a chat, so it can also get things wrong on a bigger scale: in one task it may change dozens of files. The diagram is the same as in the first lesson: you describe, AI writes, you check. The only difference is the size of the step. The stronger the tool, the more your check at the end of the loop matters. Spoiler: context, bugs and checking are the next lessons. Practice in the agent simulator.
Check yourself
- How is an agent different from a regular AI chat?
Correct answer: It reads and edits files and runs commands itself. An agent acts in the project: files, commands, tests, commits. It can still make mistakes, so checking stays your job. - The agent asks to run a command you don't understand. What do you do?
Correct answer: Ask what it does, then decide. Approving commands is your control point. A command can delete files or touch keys, so first understand what it does. - Which agent should you start with?
Correct answer: One you can access: the skill carries over. Agents are similar: the loop, branches, approving commands. The main skill, setting tasks and checking, works with any of them.